Efficient gossip and robust distributed computation

Chryssis Georgiou, Dariusz R. Kowalski, Alex A. Shvartsman

Research output: Contribution to journalArticle

22 Scopus citations

Abstract

This paper presents an efficient deterministic gossip algorithm for p synchronous, crash-prone, message-passing processors. The algorithm has time complexity T = O(log2 p) and message complexity M = O(p1+ε), for any ε > 0. This substantially improves the message complexity of the previous best algorithm that has M = O(p1.77), while maintaining the same time complexity. The strength of the new algorithm is demonstrated by constructing a deterministic algorithm for performing n tasks in this distributed setting. Previous solutions used coordinator or check-pointing approaches, immediately incurring a work penalty Ω(n + f·p) for f crashes, or relied on strong communication primitives, such as reliable broadcast, or had work too close to the trivial Θ(p·n) bound of oblivious algorithms. The new algorithm uses p crash-prone processors to perform n similar and idempotent tasks so long as one processor remains active. The work of the algorithm is W = O(n + p·min{f + 1, log3 p}) and its message complexity is M = O(fpε + pmin{f + 1, log p}), for any ε > 0. This substantially improves the work complexity of previous solutions using simple point-to-point messaging, while "meeting or beating" the corresponding message complexity bounds. The new algorithms use communication graphs and permutations with certain combinatorial properties that are shown to exist. The algorithms are correct for any permutations, and in particular, the same expected bounds can be achieved using random permutations.

Original languageEnglish (US)
Pages (from-to)224-238
Number of pages37
JournalLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume2848
Issue number1-2
DOIs
Publication statusPublished - Dec 1 2003
Externally publishedYes

    Fingerprint

Keywords

  • Combinatorial tools
  • Distributed algorithms
  • Gossip
  • Performing work
  • Processor failures

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Cite this